vgsatorras/hybrid-inference

hybrid-inference's Issues

Any reference on equation (5) of the paper?

Hi,
Just be curious on the Equation (5) in your paper, which differs from maximum value of the hidden variable in a Hidden Markov model or in a linear dynamic model.
Is there any reference that defines the most likely state of x as a linear function of the derivative of log-likelihood?

Additionally, in the loss function defined in (18), gt seems to be the ground truth of GNN variable h, which does not seem to be available since there would only exist groud truth of x and y.

Your explanation would be appreciated. Thanks.

A discussion towards better understanding of the 1st task

Dear Authors

Thanks for your great work. I am a big fun, but I have few questions towards the big picture of your work.

May I ask: Does the task of exp1_linear.py aim at recovering true trajectory positions from noisy measurements ? I notice the loss is calculated from MSE between positions and meas as in
outputs = net([operators, meas], x0, args.K, ts=ts)
mse = F.mse_loss(outputs[-1], position)

If so, Does this means that, for hybrid model, we introduce Graphical Model Message (GMM) and Graph Neural Network (GNN) mechanism to better model the unobserved variable x, for the purpose that x can flexibly take non-gaussian noise influence into account ?

Thanks for your time !!